20.01.2013 Views

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

lution face recognition [36] for Yale, PIE and ORL databases based on the successful<br />

subbands determined on testing set. From the results given in Table 3.7, we can<br />

observe that the process <strong>of</strong> decision fusion based on sum and product rules perform<br />

the best among all decision combination strategies proposed in [36]. However, in case<br />

<strong>of</strong> ORL, this strategy fails to improve upon that provided by the gray-level image<br />

using PCA.<br />

In the second case <strong>of</strong> experimentation, the successful subbands are selected using<br />

the validation set, as used in our proposed technique (see Table 3.1). Results are<br />

given in Table 3.8-3.9, in a similar manner, as in Table 3.6-3.7. We can notice that<br />

the over-tuning done in the first case <strong>of</strong>fers apparently better results, by comparing<br />

the performances given in Table 3.7 (first case) and the corresponding top seven rows<br />

<strong>of</strong> Table 3.9 (second case).<br />

Table 3.6: Best performing and successful subbands for Ekenel’s multiresolution face<br />

recognition [36] for Yale, PIE and ORL databases determined on testing set.<br />

Yale PIE ORL<br />

Best Performing Subband V3 HA2 A2<br />

A1, V2, V3 A1, H1, V1, D1 A1, A2, A3<br />

Successful Subbands H2, V2, D2<br />

Selected for Data, H3, V3, D3<br />

Feature, Decision HA3, HH3, HV3, HD3<br />

Fusion V A3, V H3, V V3, V D3<br />

DA3, DH3, DV3, DD3<br />

Hence, to provide a meaningful comparative study, which is also practically use-<br />

ful, we use the second approach (based on validation set) to compare the performance<br />

<strong>of</strong> our underlying subband face representation with that suggested in [36], using data,<br />

feature and decision level fusion. We do so with identical feature extraction and dis-<br />

tance metric, which is PCA and L2 norm in both cases. Image size is also kept as<br />

same for both cases. We use the same validation set to select the subbands performing<br />

equally good and better than original image for multiresolution face recognition. The<br />

66

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!